【摘要】 验证码是众多互联网应用的重要安全措施,但在面对基于深度学习技术的破解工具时已难以保证其安全性。文章将对抗样本与文本验证码相结合,提出基于区域更新的模型集成白盒生成算法。为提高对抗验证码的迁移性,降低未知模型的识别率,根据多个模型的预测结果对损失函数进行加权求和;在目标函数中添加扰动项,通过梯度下降的方式更新验证码图像的像素值以最小化目标函数;将验证码文本区域和背景区域分区更新,以降低对抗样本所需的扰动量同时避免识别模型预处理对扰动的破坏。实验结果表明,在面对基于深度学习的识别模型时,文章提出的算法具有更低的识别率。更多还原

【Abstract】 CAPTCHA is an important security measure for many Internet applications,it is facing many chalanges while facing cracking tools based on deep learning technology.This paper proposes a model integrated white box generation algorithm based on region update by combining the adversarial samples with text CAPTCHA.In order to improve the transferability of adversarial CAPTCHA and reduce the recognition rate of unknown models,perturbation term is added to the objective function and the pixel value of CAPTCHA image is updated by means of gradient descent to minimize the objective function.The CAPTCHA text area and background area are partitioned and updated seperately to reduce the perturbations needed to generate adversarial samples and avoid the damage to the perturbations caused by the pre-processing of crack models.The experimental results show that the proposed algorithm has a lower recognition rate than the relevant work under the condition of considerable perturbation when facing the recognition model based on deep learning,which effectively increases the security of text CAPTCHA.更多还原